xref: /csrg-svn/sys/miscfs/union/union_subr.c (revision 67401)
1 /*
2  * Copyright (c) 1994 Jan-Simon Pendry
3  * Copyright (c) 1994
4  *	The Regents of the University of California.  All rights reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * Jan-Simon Pendry.
8  *
9  * %sccs.include.redist.c%
10  *
11  *	@(#)union_subr.c	8.10 (Berkeley) 06/16/94
12  */
13 
14 #include <sys/param.h>
15 #include <sys/systm.h>
16 #include <sys/time.h>
17 #include <sys/kernel.h>
18 #include <sys/vnode.h>
19 #include <sys/namei.h>
20 #include <sys/malloc.h>
21 #include <sys/file.h>
22 #include <sys/filedesc.h>
23 #include <sys/queue.h>
24 #include <sys/mount.h>
25 #include <vm/vm.h>		/* for vnode_pager_setsize */
26 #include <miscfs/union/union.h>
27 
28 #ifdef DIAGNOSTIC
29 #include <sys/proc.h>
30 #endif
31 
32 /* must be power of two, otherwise change UNION_HASH() */
33 #define NHASH 32
34 
35 /* unsigned int ... */
36 #define UNION_HASH(u, l) \
37 	(((((unsigned long) (u)) + ((unsigned long) l)) >> 8) & (NHASH-1))
38 
39 static LIST_HEAD(unhead, union_node) unhead[NHASH];
40 static int unvplock[NHASH];
41 
42 int
43 union_init()
44 {
45 	int i;
46 
47 	for (i = 0; i < NHASH; i++)
48 		LIST_INIT(&unhead[i]);
49 	bzero((caddr_t) unvplock, sizeof(unvplock));
50 }
51 
52 static int
53 union_list_lock(ix)
54 	int ix;
55 {
56 
57 	if (unvplock[ix] & UN_LOCKED) {
58 		unvplock[ix] |= UN_WANT;
59 		sleep((caddr_t) &unvplock[ix], PINOD);
60 		return (1);
61 	}
62 
63 	unvplock[ix] |= UN_LOCKED;
64 
65 	return (0);
66 }
67 
68 static void
69 union_list_unlock(ix)
70 	int ix;
71 {
72 
73 	unvplock[ix] &= ~UN_LOCKED;
74 
75 	if (unvplock[ix] & UN_WANT) {
76 		unvplock[ix] &= ~UN_WANT;
77 		wakeup((caddr_t) &unvplock[ix]);
78 	}
79 }
80 
81 void
82 union_updatevp(un, uppervp, lowervp)
83 	struct union_node *un;
84 	struct vnode *uppervp;
85 	struct vnode *lowervp;
86 {
87 	int ohash = UNION_HASH(un->un_uppervp, un->un_lowervp);
88 	int nhash = UNION_HASH(uppervp, lowervp);
89 	int docache = (lowervp != NULLVP || uppervp != NULLVP);
90 
91 	/*
92 	 * Ensure locking is ordered from lower to higher
93 	 * to avoid deadlocks.
94 	 */
95 	if (nhash < ohash) {
96 		int t = ohash;
97 		ohash = nhash;
98 		nhash = t;
99 	}
100 
101 	if (ohash != nhash)
102 		while (union_list_lock(ohash))
103 			continue;
104 
105 	while (union_list_lock(nhash))
106 		continue;
107 
108 	if (ohash != nhash || !docache) {
109 		if (un->un_flags & UN_CACHED) {
110 			un->un_flags &= ~UN_CACHED;
111 			LIST_REMOVE(un, un_cache);
112 		}
113 	}
114 
115 	if (ohash != nhash)
116 		union_list_unlock(ohash);
117 
118 	if (un->un_lowervp != lowervp) {
119 		if (un->un_lowervp) {
120 			vrele(un->un_lowervp);
121 			if (un->un_path) {
122 				free(un->un_path, M_TEMP);
123 				un->un_path = 0;
124 			}
125 			if (un->un_dirvp) {
126 				vrele(un->un_dirvp);
127 				un->un_dirvp = NULLVP;
128 			}
129 		}
130 		un->un_lowervp = lowervp;
131 		un->un_lowersz = VNOVAL;
132 	}
133 
134 	if (un->un_uppervp != uppervp) {
135 		if (un->un_uppervp)
136 			vrele(un->un_uppervp);
137 
138 		un->un_uppervp = uppervp;
139 		un->un_uppersz = VNOVAL;
140 	}
141 
142 	if (docache && (ohash != nhash)) {
143 		LIST_INSERT_HEAD(&unhead[nhash], un, un_cache);
144 		un->un_flags |= UN_CACHED;
145 	}
146 
147 	union_list_unlock(nhash);
148 }
149 
150 void
151 union_newlower(un, lowervp)
152 	struct union_node *un;
153 	struct vnode *lowervp;
154 {
155 
156 	union_updatevp(un, un->un_uppervp, lowervp);
157 }
158 
159 void
160 union_newupper(un, uppervp)
161 	struct union_node *un;
162 	struct vnode *uppervp;
163 {
164 
165 	union_updatevp(un, uppervp, un->un_lowervp);
166 }
167 
168 /*
169  * Keep track of size changes in the underlying vnodes.
170  * If the size changes, then callback to the vm layer
171  * giving priority to the upper layer size.
172  */
173 void
174 union_newsize(vp, uppersz, lowersz)
175 	struct vnode *vp;
176 	off_t uppersz, lowersz;
177 {
178 	struct union_node *un;
179 	off_t sz;
180 
181 	/* only interested in regular files */
182 	if (vp->v_type != VREG)
183 		return;
184 
185 	un = VTOUNION(vp);
186 	sz = VNOVAL;
187 
188 	if ((uppersz != VNOVAL) && (un->un_uppersz != uppersz)) {
189 		un->un_uppersz = uppersz;
190 		if (sz == VNOVAL)
191 			sz = un->un_uppersz;
192 	}
193 
194 	if ((lowersz != VNOVAL) && (un->un_lowersz != lowersz)) {
195 		un->un_lowersz = lowersz;
196 		if (sz == VNOVAL)
197 			sz = un->un_lowersz;
198 	}
199 
200 	if (sz != VNOVAL) {
201 #ifdef UNION_DIAGNOSTIC
202 		printf("union: %s size now %ld\n",
203 			uppersz != VNOVAL ? "upper" : "lower", (long) sz);
204 #endif
205 		vnode_pager_setsize(vp, sz);
206 	}
207 }
208 
209 /*
210  * allocate a union_node/vnode pair.  the vnode is
211  * referenced and locked.  the new vnode is returned
212  * via (vpp).  (mp) is the mountpoint of the union filesystem,
213  * (dvp) is the parent directory where the upper layer object
214  * should exist (but doesn't) and (cnp) is the componentname
215  * information which is partially copied to allow the upper
216  * layer object to be created at a later time.  (uppervp)
217  * and (lowervp) reference the upper and lower layer objects
218  * being mapped.  either, but not both, can be nil.
219  * if supplied, (uppervp) is locked.
220  * the reference is either maintained in the new union_node
221  * object which is allocated, or they are vrele'd.
222  *
223  * all union_nodes are maintained on a singly-linked
224  * list.  new nodes are only allocated when they cannot
225  * be found on this list.  entries on the list are
226  * removed when the vfs reclaim entry is called.
227  *
228  * a single lock is kept for the entire list.  this is
229  * needed because the getnewvnode() function can block
230  * waiting for a vnode to become free, in which case there
231  * may be more than one process trying to get the same
232  * vnode.  this lock is only taken if we are going to
233  * call getnewvnode, since the kernel itself is single-threaded.
234  *
235  * if an entry is found on the list, then call vget() to
236  * take a reference.  this is done because there may be
237  * zero references to it and so it needs to removed from
238  * the vnode free list.
239  */
240 int
241 union_allocvp(vpp, mp, undvp, dvp, cnp, uppervp, lowervp)
242 	struct vnode **vpp;
243 	struct mount *mp;
244 	struct vnode *undvp;
245 	struct vnode *dvp;		/* may be null */
246 	struct componentname *cnp;	/* may be null */
247 	struct vnode *uppervp;		/* may be null */
248 	struct vnode *lowervp;		/* may be null */
249 {
250 	int error;
251 	struct union_node *un;
252 	struct union_node **pp;
253 	struct vnode *xlowervp = NULLVP;
254 	struct union_mount *um = MOUNTTOUNIONMOUNT(mp);
255 	int hash;
256 	int vflag;
257 	int try;
258 
259 	if (uppervp == NULLVP && lowervp == NULLVP)
260 		panic("union: unidentifiable allocation");
261 
262 	if (uppervp && lowervp && (uppervp->v_type != lowervp->v_type)) {
263 		xlowervp = lowervp;
264 		lowervp = NULLVP;
265 	}
266 
267 	/* detect the root vnode (and aliases) */
268 	vflag = 0;
269 	if ((uppervp == um->um_uppervp) &&
270 	    ((lowervp == NULLVP) || lowervp == um->um_lowervp)) {
271 		if (lowervp == NULLVP) {
272 			lowervp = um->um_lowervp;
273 			if (lowervp != NULLVP)
274 				VREF(lowervp);
275 		}
276 		vflag = VROOT;
277 	}
278 
279 loop:
280 	for (try = 0; try < 3; try++) {
281 		switch (try) {
282 		case 0:
283 			if (lowervp == NULLVP)
284 				continue;
285 			hash = UNION_HASH(uppervp, lowervp);
286 			break;
287 
288 		case 1:
289 			if (uppervp == NULLVP)
290 				continue;
291 			hash = UNION_HASH(uppervp, NULLVP);
292 			break;
293 
294 		case 2:
295 			if (lowervp == NULLVP)
296 				continue;
297 			hash = UNION_HASH(NULLVP, lowervp);
298 			break;
299 		}
300 
301 		while (union_list_lock(hash))
302 			continue;
303 
304 		for (un = unhead[hash].lh_first; un != 0;
305 					un = un->un_cache.le_next) {
306 			if ((un->un_lowervp == lowervp ||
307 			     un->un_lowervp == NULLVP) &&
308 			    (un->un_uppervp == uppervp ||
309 			     un->un_uppervp == NULLVP) &&
310 			    (UNIONTOV(un)->v_mount == mp)) {
311 				if (vget(UNIONTOV(un), 0)) {
312 					union_list_unlock(hash);
313 					goto loop;
314 				}
315 				break;
316 			}
317 		}
318 
319 		union_list_unlock(hash);
320 
321 		if (un)
322 			break;
323 	}
324 
325 	if (un) {
326 		/*
327 		 * Obtain a lock on the union_node.
328 		 * uppervp is locked, though un->un_uppervp
329 		 * may not be.  this doesn't break the locking
330 		 * hierarchy since in the case that un->un_uppervp
331 		 * is not yet locked it will be vrele'd and replaced
332 		 * with uppervp.
333 		 */
334 
335 		if ((dvp != NULLVP) && (uppervp == dvp)) {
336 			/*
337 			 * Access ``.'', so (un) will already
338 			 * be locked.  Since this process has
339 			 * the lock on (uppervp) no other
340 			 * process can hold the lock on (un).
341 			 */
342 #ifdef DIAGNOSTIC
343 			if ((un->un_flags & UN_LOCKED) == 0)
344 				panic("union: . not locked");
345 			else if (curproc && un->un_pid != curproc->p_pid &&
346 				    un->un_pid > -1 && curproc->p_pid > -1)
347 				panic("union: allocvp not lock owner");
348 #endif
349 		} else {
350 			if (un->un_flags & UN_LOCKED) {
351 				vrele(UNIONTOV(un));
352 				un->un_flags |= UN_WANT;
353 				sleep((caddr_t) &un->un_flags, PINOD);
354 				goto loop;
355 			}
356 			un->un_flags |= UN_LOCKED;
357 
358 #ifdef DIAGNOSTIC
359 			if (curproc)
360 				un->un_pid = curproc->p_pid;
361 			else
362 				un->un_pid = -1;
363 #endif
364 		}
365 
366 		/*
367 		 * At this point, the union_node is locked,
368 		 * un->un_uppervp may not be locked, and uppervp
369 		 * is locked or nil.
370 		 */
371 
372 		/*
373 		 * Save information about the upper layer.
374 		 */
375 		if (uppervp != un->un_uppervp) {
376 			union_newupper(un, uppervp);
377 		} else if (uppervp) {
378 			vrele(uppervp);
379 		}
380 
381 		if (un->un_uppervp) {
382 			un->un_flags |= UN_ULOCK;
383 			un->un_flags &= ~UN_KLOCK;
384 		}
385 
386 		/*
387 		 * Save information about the lower layer.
388 		 * This needs to keep track of pathname
389 		 * and directory information which union_vn_create
390 		 * might need.
391 		 */
392 		if (lowervp != un->un_lowervp) {
393 			union_newlower(un, lowervp);
394 			if (cnp && (lowervp != NULLVP) &&
395 			    (lowervp->v_type == VREG)) {
396 				un->un_hash = cnp->cn_hash;
397 				un->un_path = malloc(cnp->cn_namelen+1,
398 						M_TEMP, M_WAITOK);
399 				bcopy(cnp->cn_nameptr, un->un_path,
400 						cnp->cn_namelen);
401 				un->un_path[cnp->cn_namelen] = '\0';
402 				VREF(dvp);
403 				un->un_dirvp = dvp;
404 			}
405 		} else if (lowervp) {
406 			vrele(lowervp);
407 		}
408 		*vpp = UNIONTOV(un);
409 		return (0);
410 	}
411 
412 	/*
413 	 * otherwise lock the vp list while we call getnewvnode
414 	 * since that can block.
415 	 */
416 	hash = UNION_HASH(uppervp, lowervp);
417 
418 	if (union_list_lock(hash))
419 		goto loop;
420 
421 	error = getnewvnode(VT_UNION, mp, union_vnodeop_p, vpp);
422 	if (error) {
423 		if (uppervp) {
424 			if (dvp == uppervp)
425 				vrele(uppervp);
426 			else
427 				vput(uppervp);
428 		}
429 		if (lowervp)
430 			vrele(lowervp);
431 
432 		goto out;
433 	}
434 
435 	MALLOC((*vpp)->v_data, void *, sizeof(struct union_node),
436 		M_TEMP, M_WAITOK);
437 
438 	(*vpp)->v_flag |= vflag;
439 	if (uppervp)
440 		(*vpp)->v_type = uppervp->v_type;
441 	else
442 		(*vpp)->v_type = lowervp->v_type;
443 	un = VTOUNION(*vpp);
444 	un->un_vnode = *vpp;
445 	un->un_uppervp = uppervp;
446 	un->un_uppersz = VNOVAL;
447 	un->un_lowervp = lowervp;
448 	un->un_lowersz = VNOVAL;
449 	un->un_openl = 0;
450 	un->un_flags = UN_LOCKED;
451 	if (un->un_uppervp)
452 		un->un_flags |= UN_ULOCK;
453 #ifdef DIAGNOSTIC
454 	if (curproc)
455 		un->un_pid = curproc->p_pid;
456 	else
457 		un->un_pid = -1;
458 #endif
459 	if (cnp && (lowervp != NULLVP) && (lowervp->v_type == VREG)) {
460 		un->un_hash = cnp->cn_hash;
461 		un->un_path = malloc(cnp->cn_namelen+1, M_TEMP, M_WAITOK);
462 		bcopy(cnp->cn_nameptr, un->un_path, cnp->cn_namelen);
463 		un->un_path[cnp->cn_namelen] = '\0';
464 		VREF(dvp);
465 		un->un_dirvp = dvp;
466 	} else {
467 		un->un_hash = 0;
468 		un->un_path = 0;
469 		un->un_dirvp = 0;
470 	}
471 
472 	LIST_INSERT_HEAD(&unhead[hash], un, un_cache);
473 	un->un_flags |= UN_CACHED;
474 
475 	if (xlowervp)
476 		vrele(xlowervp);
477 
478 out:
479 	union_list_unlock(hash);
480 
481 	return (error);
482 }
483 
484 int
485 union_freevp(vp)
486 	struct vnode *vp;
487 {
488 	struct union_node *un = VTOUNION(vp);
489 
490 	if (un->un_flags & UN_CACHED) {
491 		un->un_flags &= ~UN_CACHED;
492 		LIST_REMOVE(un, un_cache);
493 	}
494 
495 	if (un->un_uppervp != NULLVP)
496 		vrele(un->un_uppervp);
497 	if (un->un_lowervp != NULLVP)
498 		vrele(un->un_lowervp);
499 	if (un->un_dirvp != NULLVP)
500 		vrele(un->un_dirvp);
501 	if (un->un_path)
502 		free(un->un_path, M_TEMP);
503 
504 	FREE(vp->v_data, M_TEMP);
505 	vp->v_data = 0;
506 
507 	return (0);
508 }
509 
510 /*
511  * copyfile.  copy the vnode (fvp) to the vnode (tvp)
512  * using a sequence of reads and writes.  both (fvp)
513  * and (tvp) are locked on entry and exit.
514  */
515 int
516 union_copyfile(fvp, tvp, cred, p)
517 	struct vnode *fvp;
518 	struct vnode *tvp;
519 	struct ucred *cred;
520 	struct proc *p;
521 {
522 	char *buf;
523 	struct uio uio;
524 	struct iovec iov;
525 	int error = 0;
526 
527 	/*
528 	 * strategy:
529 	 * allocate a buffer of size MAXBSIZE.
530 	 * loop doing reads and writes, keeping track
531 	 * of the current uio offset.
532 	 * give up at the first sign of trouble.
533 	 */
534 
535 	uio.uio_procp = p;
536 	uio.uio_segflg = UIO_SYSSPACE;
537 	uio.uio_offset = 0;
538 
539 	VOP_UNLOCK(fvp);				/* XXX */
540 	LEASE_CHECK(fvp, p, cred, LEASE_READ);
541 	VOP_LOCK(fvp);					/* XXX */
542 	VOP_UNLOCK(tvp);				/* XXX */
543 	LEASE_CHECK(tvp, p, cred, LEASE_WRITE);
544 	VOP_LOCK(tvp);					/* XXX */
545 
546 	buf = malloc(MAXBSIZE, M_TEMP, M_WAITOK);
547 
548 	/* ugly loop follows... */
549 	do {
550 		off_t offset = uio.uio_offset;
551 
552 		uio.uio_iov = &iov;
553 		uio.uio_iovcnt = 1;
554 		iov.iov_base = buf;
555 		iov.iov_len = MAXBSIZE;
556 		uio.uio_resid = iov.iov_len;
557 		uio.uio_rw = UIO_READ;
558 		error = VOP_READ(fvp, &uio, 0, cred);
559 
560 		if (error == 0) {
561 			uio.uio_iov = &iov;
562 			uio.uio_iovcnt = 1;
563 			iov.iov_base = buf;
564 			iov.iov_len = MAXBSIZE - uio.uio_resid;
565 			uio.uio_offset = offset;
566 			uio.uio_rw = UIO_WRITE;
567 			uio.uio_resid = iov.iov_len;
568 
569 			if (uio.uio_resid == 0)
570 				break;
571 
572 			do {
573 				error = VOP_WRITE(tvp, &uio, 0, cred);
574 			} while ((uio.uio_resid > 0) && (error == 0));
575 		}
576 
577 	} while (error == 0);
578 
579 	free(buf, M_TEMP);
580 	return (error);
581 }
582 
583 /*
584  * (un) is assumed to be locked on entry and remains
585  * locked on exit.
586  */
587 int
588 union_copyup(un, docopy, cred, p)
589 	struct union_node *un;
590 	int docopy;
591 	struct ucred *cred;
592 	struct proc *p;
593 {
594 	int error;
595 	struct vnode *lvp, *uvp;
596 
597 	error = union_vn_create(&uvp, un, p);
598 	if (error)
599 		return (error);
600 
601 	/* at this point, uppervp is locked */
602 	union_newupper(un, uvp);
603 	un->un_flags |= UN_ULOCK;
604 
605 	lvp = un->un_lowervp;
606 
607 	if (docopy) {
608 		/*
609 		 * XX - should not ignore errors
610 		 * from VOP_CLOSE
611 		 */
612 		VOP_LOCK(lvp);
613 		error = VOP_OPEN(lvp, FREAD, cred, p);
614 		if (error == 0) {
615 			error = union_copyfile(lvp, uvp, cred, p);
616 			VOP_UNLOCK(lvp);
617 			(void) VOP_CLOSE(lvp, FREAD);
618 		}
619 #ifdef UNION_DIAGNOSTIC
620 		if (error == 0)
621 			uprintf("union: copied up %s\n", un->un_path);
622 #endif
623 
624 	}
625 	un->un_flags &= ~UN_ULOCK;
626 	VOP_UNLOCK(uvp);
627 	union_vn_close(uvp, FWRITE, cred, p);
628 	VOP_LOCK(uvp);
629 	un->un_flags |= UN_ULOCK;
630 
631 	/*
632 	 * Subsequent IOs will go to the top layer, so
633 	 * call close on the lower vnode and open on the
634 	 * upper vnode to ensure that the filesystem keeps
635 	 * its references counts right.  This doesn't do
636 	 * the right thing with (cred) and (FREAD) though.
637 	 * Ignoring error returns is not right, either.
638 	 */
639 	if (error == 0) {
640 		int i;
641 
642 		for (i = 0; i < un->un_openl; i++) {
643 			(void) VOP_CLOSE(lvp, FREAD);
644 			(void) VOP_OPEN(uvp, FREAD, cred, p);
645 		}
646 		un->un_openl = 0;
647 	}
648 
649 	return (error);
650 
651 }
652 
653 /*
654  * Create a shadow directory in the upper layer.
655  * The new vnode is returned locked.
656  *
657  * (um) points to the union mount structure for access to the
658  * the mounting process's credentials.
659  * (dvp) is the directory in which to create the shadow directory.
660  * it is unlocked on entry and exit.
661  * (cnp) is the componentname to be created.
662  * (vpp) is the returned newly created shadow directory, which
663  * is returned locked.
664  */
665 int
666 union_mkshadow(um, dvp, cnp, vpp)
667 	struct union_mount *um;
668 	struct vnode *dvp;
669 	struct componentname *cnp;
670 	struct vnode **vpp;
671 {
672 	int error;
673 	struct vattr va;
674 	struct proc *p = cnp->cn_proc;
675 	struct componentname cn;
676 
677 	/*
678 	 * policy: when creating the shadow directory in the
679 	 * upper layer, create it owned by the user who did
680 	 * the mount, group from parent directory, and mode
681 	 * 777 modified by umask (ie mostly identical to the
682 	 * mkdir syscall).  (jsp, kb)
683 	 */
684 
685 	/*
686 	 * A new componentname structure must be faked up because
687 	 * there is no way to know where the upper level cnp came
688 	 * from or what it is being used for.  This must duplicate
689 	 * some of the work done by NDINIT, some of the work done
690 	 * by namei, some of the work done by lookup and some of
691 	 * the work done by VOP_LOOKUP when given a CREATE flag.
692 	 * Conclusion: Horrible.
693 	 *
694 	 * The pathname buffer will be FREEed by VOP_MKDIR.
695 	 */
696 	cn.cn_pnbuf = malloc(cnp->cn_namelen+1, M_NAMEI, M_WAITOK);
697 	bcopy(cnp->cn_nameptr, cn.cn_pnbuf, cnp->cn_namelen);
698 	cn.cn_pnbuf[cnp->cn_namelen] = '\0';
699 
700 	cn.cn_nameiop = CREATE;
701 	cn.cn_flags = (LOCKPARENT|HASBUF|SAVENAME|SAVESTART|ISLASTCN);
702 	cn.cn_proc = cnp->cn_proc;
703 	if (um->um_op == UNMNT_ABOVE)
704 		cn.cn_cred = cnp->cn_cred;
705 	else
706 		cn.cn_cred = um->um_cred;
707 	cn.cn_nameptr = cn.cn_pnbuf;
708 	cn.cn_namelen = cnp->cn_namelen;
709 	cn.cn_hash = cnp->cn_hash;
710 	cn.cn_consume = cnp->cn_consume;
711 
712 	VREF(dvp);
713 	if (error = relookup(dvp, vpp, &cn))
714 		return (error);
715 	vrele(dvp);
716 
717 	if (*vpp) {
718 		VOP_ABORTOP(dvp, &cn);
719 		VOP_UNLOCK(dvp);
720 		vrele(*vpp);
721 		*vpp = NULLVP;
722 		return (EEXIST);
723 	}
724 
725 	VATTR_NULL(&va);
726 	va.va_type = VDIR;
727 	va.va_mode = um->um_cmode;
728 
729 	/* LEASE_CHECK: dvp is locked */
730 	LEASE_CHECK(dvp, p, p->p_ucred, LEASE_WRITE);
731 
732 	error = VOP_MKDIR(dvp, vpp, &cn, &va);
733 	return (error);
734 }
735 
736 /*
737  * union_vn_create: creates and opens a new shadow file
738  * on the upper union layer.  this function is similar
739  * in spirit to calling vn_open but it avoids calling namei().
740  * the problem with calling namei is that a) it locks too many
741  * things, and b) it doesn't start at the "right" directory,
742  * whereas relookup is told where to start.
743  */
744 int
745 union_vn_create(vpp, un, p)
746 	struct vnode **vpp;
747 	struct union_node *un;
748 	struct proc *p;
749 {
750 	struct vnode *vp;
751 	struct ucred *cred = p->p_ucred;
752 	struct vattr vat;
753 	struct vattr *vap = &vat;
754 	int fmode = FFLAGS(O_WRONLY|O_CREAT|O_TRUNC|O_EXCL);
755 	int error;
756 	int cmode = UN_FILEMODE & ~p->p_fd->fd_cmask;
757 	char *cp;
758 	struct componentname cn;
759 
760 	*vpp = NULLVP;
761 
762 	/*
763 	 * Build a new componentname structure (for the same
764 	 * reasons outlines in union_mkshadow).
765 	 * The difference here is that the file is owned by
766 	 * the current user, rather than by the person who
767 	 * did the mount, since the current user needs to be
768 	 * able to write the file (that's why it is being
769 	 * copied in the first place).
770 	 */
771 	cn.cn_namelen = strlen(un->un_path);
772 	cn.cn_pnbuf = (caddr_t) malloc(cn.cn_namelen, M_NAMEI, M_WAITOK);
773 	bcopy(un->un_path, cn.cn_pnbuf, cn.cn_namelen+1);
774 	cn.cn_nameiop = CREATE;
775 	cn.cn_flags = (LOCKPARENT|HASBUF|SAVENAME|SAVESTART|ISLASTCN);
776 	cn.cn_proc = p;
777 	cn.cn_cred = p->p_ucred;
778 	cn.cn_nameptr = cn.cn_pnbuf;
779 	cn.cn_hash = un->un_hash;
780 	cn.cn_consume = 0;
781 
782 	VREF(un->un_dirvp);
783 	if (error = relookup(un->un_dirvp, &vp, &cn))
784 		return (error);
785 	vrele(un->un_dirvp);
786 
787 	if (vp) {
788 		VOP_ABORTOP(un->un_dirvp, &cn);
789 		if (un->un_dirvp == vp)
790 			vrele(un->un_dirvp);
791 		else
792 			vput(un->un_dirvp);
793 		vrele(vp);
794 		return (EEXIST);
795 	}
796 
797 	/*
798 	 * Good - there was no race to create the file
799 	 * so go ahead and create it.  The permissions
800 	 * on the file will be 0666 modified by the
801 	 * current user's umask.  Access to the file, while
802 	 * it is unioned, will require access to the top *and*
803 	 * bottom files.  Access when not unioned will simply
804 	 * require access to the top-level file.
805 	 * TODO: confirm choice of access permissions.
806 	 */
807 	VATTR_NULL(vap);
808 	vap->va_type = VREG;
809 	vap->va_mode = cmode;
810 	LEASE_CHECK(un->un_dirvp, p, cred, LEASE_WRITE);
811 	if (error = VOP_CREATE(un->un_dirvp, &vp, &cn, vap))
812 		return (error);
813 
814 	if (error = VOP_OPEN(vp, fmode, cred, p)) {
815 		vput(vp);
816 		return (error);
817 	}
818 
819 	vp->v_writecount++;
820 	*vpp = vp;
821 	return (0);
822 }
823 
824 int
825 union_vn_close(vp, fmode, cred, p)
826 	struct vnode *vp;
827 	int fmode;
828 	struct ucred *cred;
829 	struct proc *p;
830 {
831 	if (fmode & FWRITE)
832 		--vp->v_writecount;
833 	return (VOP_CLOSE(vp, fmode));
834 }
835 
836 void
837 union_removed_upper(un)
838 	struct union_node *un;
839 {
840 
841 	if (un->un_flags & UN_ULOCK) {
842 		un->un_flags &= ~UN_ULOCK;
843 		VOP_UNLOCK(un->un_uppervp);
844 	}
845 
846 	if (un->un_flags & UN_CACHED) {
847 		un->un_flags &= ~UN_CACHED;
848 		LIST_REMOVE(un, un_cache);
849 	}
850 }
851 
852 struct vnode *
853 union_lowervp(vp)
854 	struct vnode *vp;
855 {
856 	struct union_node *un = VTOUNION(vp);
857 
858 	if ((un->un_lowervp != NULLVP) &&
859 	    (vp->v_type == un->un_lowervp->v_type)) {
860 		if (vget(un->un_lowervp, 0) == 0)
861 			return (un->un_lowervp);
862 	}
863 
864 	return (NULLVP);
865 }
866